Daily Stock Trend Forecasting using Fuzzy Rule-based Reinforcement Learning Agent and Portfolio Management

Transcription

1 Daily Stock Trend Forecasting using Fuzzy Rule-based Reinforcement Learning Agent and Portfolio Management Rubell Marion Lincy G. Jessy John C. Department of Mathematics School of Natural Sciences National Institute of Technology Calicut Calicut, Kerala India. ubell Marion Lincy G., Jessy John C. (NIT) International Symposium on Forecasting 1 / 20

2 Outline 1 Motivation 2 Proposed System Portfolio Selection and Data Pre-processing Fuzzy Rule-based System Daily Stock Trend Forecasting Portfolio Construction 3 Experimental Results 4 Conclusion & Future Work 5 References Rubell Marion Lincy G., Jessy John C. (NIT) International Symposium on Forecasting 2 / 20

3 Outline 1 Motivation 2 Proposed System Portfolio Selection and Data Pre-processing Fuzzy Rule-based System Daily Stock Trend Forecasting Portfolio Construction 3 Experimental Results 4 Conclusion & Future Work 5 References Rubell Marion Lincy G., Jessy John C. (NIT) International Symposium on Forecasting 3 / 20

4 Motivation Many of the research works in the area of financial trading aims at developing an automated agent for stock trading in different scenarios [1, 2, 3] The concept of fuzzy mathematics is used in many ways for developing a portfolio [4, 5, 6] Inspired by these works, we have developed daily stock trend forecasting agent which determines whether to Buy/Sell/Hold a particular stock on daily basis which in the latter part of the system used in selecting the appropriate stocks for developing a portfolio Rubell Marion Lincy G., Jessy John C. (NIT) International Symposium on Forecasting 4 / 20

5 Outline 1 Motivation 2 Proposed System Portfolio Selection and Data Pre-processing Fuzzy Rule-based System Daily Stock Trend Forecasting Portfolio Construction 3 Experimental Results 4 Conclusion & Future Work 5 References Rubell Marion Lincy G., Jessy John C. (NIT) International Symposium on Forecasting 5 / 20

6 System Architecture Figure: Structure of the Proposed System Rubell Marion Lincy G., Jessy John C. (NIT) International Symposium on Forecasting 6 / 20

7 Portfolio Selection The first step in constructing a portfolio is the selection of initial set of stocks In this work we consider six sectors from the Bombay Stock Exchange (BSE) which are as follows: Banks - Private Sector Banks - Public Sector Diversified Domestic Appliances Power - Generation and Distribution Refineries Initially a total of 80 stocks are chosen The stocks with a negative Price-to-earnings ratio or Shareholder s Equity Value are eliminated Rubell Marion Lincy G., Jessy John C. (NIT) International Symposium on Forecasting 7 / 20

8 Pre-processing Both Fundamental and Technical Data of each stock is considered in the process of rating the stock as to be eligible for a portfolio Fundamental Data 1 Operating Profit 2 Profit before taxes 3 Return on Net Worth 4 Leverage 5 Price to Earnings 6 Dividend Yield 7 Price to Book Value 8 Market Capitalization Rubell Marion Lincy G., Jessy John C. (NIT) International Symposium on Forecasting 7 / 20

9 Pre-processing Technical Data 1 Average True Range 2 Bollinger Bands 3 Chaikin Oscillator 4 Moving Average Convergence Divergence 5 On Balance Volume The Relative ratio for all the above mentioned data are calculated w.r.t. each sector as follows: Relative Ratio = Particular Data Industry Average of that data Rubell Marion Lincy G., Jessy John C. (NIT) International Symposium on Forecasting 7 / 20

10 Fuzzy Rule-based System The data is processed to assign rating for each stock The rating is assigned by using a Mamdani type Fuzzy Rule-based System This method was introduced by Mamdani (1975) Mamdani s method is the most commonly used in applications, due to its simple structure of min-max operations The defuzzification process used is the largest (absolute) value of maximum The stocks are rated from 0 to 100 Rubell Marion Lincy G., Jessy John C. (NIT) International Symposium on Forecasting 8 / 20

11 Daily Stock Trend Forecasting The suggestions of whether to buy/hold/sell is determined by the daily stock trend forecasting agent Using the technical data of the stock, a trend value is determined as follows: Trend Value = (close high) + (Close Low) 2 A Sugeno type fuzzy rule-based system with reinforcement learning techniques is used to obtain an automated financial trading system which can decide on whether to buy or sell a stock or to stay out of the market (hold) in the daily stock trading environment Rubell Marion Lincy G., Jessy John C. (NIT) International Symposium on Forecasting 9 / 20

12 Daily Stock Trend Forecasting In general, a fuzzy rule-based system consists of the following steps: Step 1: Evaluate the antecedent for each rule Step 2: Obtain each rule s conclusion Step 3: Aggregate conclusions Step 4: Defuzzification In the proposed system, a Mamdani type fuzzy inference system is used to rate the stocks and the Sugeno type fuzzy inference system is used for providing suggestion on each stock on daily basis Rubell Marion Lincy G., Jessy John C. (NIT) International Symposium on Forecasting 9 / 20

13 Mamdani FIS vs. Sugeno FIS The main difference between the two methods lies in the consequent part of the fuzzy rules The most fundamental difference between Mamdani-type and Sugeno-type FIS is the way the crisp output is generated from the fuzzy inputs While Mamdani-type FIS uses the technique of defuzzification of a fuzzy output, Sugeno-type FIS uses weighted average to compute the crisp output Rubell Marion Lincy G., Jessy John C. (NIT) International Symposium on Forecasting 10 / 20

14 Mamdani FIS vs. Sugeno FIS Sugeno FIS has better processing time since the weighted average replace the time consuming defuzzification process Mamdani FIS has output membership functions whereas Sugeno FIS has no output membership functions Mamdani FIS is less flexible in system design in comparison to Sugeno FIS as latter can be integrated with ANFIS tool to optimize the outputs Rubell Marion Lincy G., Jessy John C. (NIT) International Symposium on Forecasting 10 / 20

15 Portfolio Construction The stocks that have a Buy/Hold suggestion is selected and their corresponding rating is taken into account Getting a capital amount(c) from the user, the following Integer Programming problem is modelled to determine the number of shares(x i ) of highly rated stocks to be bought with LB, UB denoting the lower and upper bound for x i, i = 1, 2,..., N(Number of Stocks) Maximize r i x i Subject to the Constraints N P i x i c i=1 x i UB x i LB and all x i s are positive integers Rubell Marion Lincy G., Jessy John C. (NIT) International Symposium on Forecasting 11 / 20

16 Outline 1 Motivation 2 Proposed System Portfolio Selection and Data Pre-processing Fuzzy Rule-based System Daily Stock Trend Forecasting Portfolio Construction 3 Experimental Results 4 Conclusion & Future Work 5 References Rubell Marion Lincy G., Jessy John C. (NIT) International Symposium on Forecasting 12 / 20

17 Experimental Results To start the process of rating the stocks, the fundamental data is considered and the portfolio starts on the stock day immediately after the Financial year end (April 2011) The portfolio is updated each year using the technical data that is available for the selection of stocks The experiment is carried out for 3 consecutive years with the portfolio renewed at the start of the financial year The Portfolio gives the following returns at the end of each year Year Capital Amount (Rs.) % of Return Rubell Marion Lincy G., Jessy John C. (NIT) International Symposium on Forecasting 13 / 20

18 Outline 1 Motivation 2 Proposed System Portfolio Selection and Data Pre-processing Fuzzy Rule-based System Daily Stock Trend Forecasting Portfolio Construction 3 Experimental Results 4 Conclusion & Future Work 5 References Rubell Marion Lincy G., Jessy John C. (NIT) International Symposium on Forecasting 14 / 20

19 Conclusion The Portfolio management for a selection of 80 stocks from six different industrial sectors is done The Returns show a profitable selection of stocks for the portfolio The portfolio renewal is done in a yearly basis In the consecutive years, this system considers only the technical data Rubell Marion Lincy G., Jessy John C. (NIT) International Symposium on Forecasting 15 / 20

20 Future Work A comparison can be made to study the returns of the portfolio by updating the portfolio every 3-months or 6-months or 9-months The system can be made to take both fundamental and technical data into account except for the initial year Rubell Marion Lincy G., Jessy John C. (NIT) International Symposium on Forecasting 16 / 20

21 Outline 1 Motivation 2 Proposed System Portfolio Selection and Data Pre-processing Fuzzy Rule-based System Daily Stock Trend Forecasting Portfolio Construction 3 Experimental Results 4 Conclusion & Future Work 5 References Rubell Marion Lincy G., Jessy John C. (NIT) International Symposium on Forecasting 17 / 20

22 References I [1] Stuart Duerson, FS Khan, Victor Kovalev, and Ali Hisham Malik. Reinforcement learning in online stock trading systems, [2] Jae Won Lee, Jonghun Park, O Jangmin, Jongwoo Lee, and Euyseok Hong. A multiagent approach to q-learning for daily stock trading. Systems, Man and Cybernetics, Part A: Systems and Humans, IEEE Transactions on, 37(6): , [3] Francesco Bertoluzzo and Marco Corazza. Reinforcement learning for automatic financial trading: Introduction and some applications. University Ca Foscari of Venice, Dept. of Economics Research Paper Series No, 33, Rubell Marion Lincy G., Jessy John C. (NIT) International Symposium on Forecasting 18 / 20

23 References II [4] Chinho Lin and Ping-Jung Hsieh. A fuzzy decision support system for strategic portfolio management. Decision Support Systems, 38(3): , [5] M Arenas Parra, Amelia Bilbao Terol, and MV Rodrıguez Urıa. A fuzzy goal programming approach to portfolio selection. European Journal of Operational Research, 133(2): , [6] Ralf Östermark. A fuzzy control model (fcm) for dynamic portfolio management. Fuzzy sets and Systems, 78(3): , Rubell Marion Lincy G., Jessy John C. (NIT) International Symposium on Forecasting 19 / 20

24 Thank You Rubell Marion Lincy G., Jessy John C. (NIT) International Symposium on Forecasting 20 / 20